Modeling the semantics of emotion: A culturally grounded approach to Flemish narratives

Abstract

Expression of emotions through natural language is deeply embedded in culturally specific contexts and extends beyond simple lexical labels. A central difficulty lies in extracting structured semantic knowledge from unstructured daily narratives, a task particularly challenging for under-resourced language varieties such as Flemish (Belgian Dutch), which have historically received minimal computational attention. This study evaluates transformer-based models, specifically BERTopic, against traditional co-occurrence-based models (LDA) and clustering baselines (KMeans) on a uniquely large corpus of 24,854 daily narratives collected from 102 Dutch speakers over 70 days in Belgium, using both automated coherence metrics and a human evaluation. Our findings demonstrate that moving beyond frequency-based models is essential for semantically and culturally accurate analysis of naturalistic emotional language in under-resourced varieties.